【電腦科學】【2018.02】自適應運動規劃

本文為美國卡內基·梅隆大學（作者：Sanjiban Choudhury）的博士論文，共255頁。

移動機器人正越來越多地被部署在現實世界中，以適應諸如運輸、配送和檢查等應用的不斷增長需求。移動機器人的運動規劃系統在它們所遇到的各種場景中都應該具有一致的效能。儘管可以使用具有可證明的最壞情況保證的最先進的規劃器來解決這些規劃問題，但是它們在有限時間內應對不同場景的效能是不同的。本文提出的機器人規劃模組必須根據所遇到的規劃問題的分佈來調整其搜尋策略，以達到實時效能。我們解決這個問題主要有三個挑戰。

首先，我們證明了即使規劃問題的分佈是固定的，由於規劃策略的效能隨著環境的微小變化而波動，因此，設計非自適應規劃器也是具有挑戰性的。我們描述了互補策略的存在性，並提出通過執行不同的規劃器集合來實現。

其次，當分佈發生變化時，我們需要一個元規劃器，它可以從黑盒規劃器庫中自動選擇相應的集合。我們的研究表明，針對失敗案例貪婪地訓練一個預測器列表，能夠實現一個有效的元規劃器。對於沒有訓練資料的情況，研究表明，我們可以通過採用線上尋呼理論中的演算法來動態地學習整合。

第三，為了提高效率，我們需要一個白盒規劃器，它在規劃週期中直接調整搜尋策略。我們在資料驅動的模仿學習框架下提出了一個有效的訓練自適應搜尋的啟發方法。我們還提出了一種創新的與貝葉斯主動學習相關的連線方法，並提出了自適應評估圖邊緣的演算法。

我們的方法最終實現了一個魯棒的實時規劃模組，該模組允許無人機無縫地跨環境跨速度導航。我們在當前框架下對一系列規劃問題進行了評估，給出了三架無人機（一架全自主無人直升機、一架大型六旋翼無人機和一架小型四旋翼無人機）平臺閉環運動的結果。雖然本論文的研究物件是移動機器人，但這些演算法可以廣泛地應用於其它問題領域，如資訊路徑規劃和操作規劃。我們還在運動規劃和主動學習、模擬學習、線上尋呼等不同領域之間建立了新穎的聯絡，為若干新的研究問題開啟了大門。

Mobile robots are increasingly beingdeployed in the real world in response to a heightened demand for applicationssuch as transportation, delivery and inspection. The motion planning systemsfor these robots are expected to have consistent performance across the widerange of scenarios that they encounter. While state-of-the-art planners, withprovable worst-case guarantees, can be employed to solve these planningproblems, their finite time performance varies across scenarios. This thesisproposes that the planning module for a robot must adapt its search strategy tothe distribution of planning problems encountered to achieve real-timeperformance. We address three principal challenges of this problem. Firstly, weshow that even when the planning problem distribution is fixed, designing anonadaptive planner can be challenging as the performance of planningstrategies fluctuates with small changes in the environment. We characterizethe existence of complementary strategies and propose to hedge our bets byexecuting a diverse ensemble of planners. Secondly, when the distribution isvarying, we require a meta-planner that can automatically select such anensemble from a library of black-box planners. We show that greedily training alist of predictors to focus on failure cases leads to an effectivemeta-planner. For situations where we have no training data, we show that wecan learn an ensemble on-the-fly by adopting algorithms from online pagingtheory. Thirdly, in the interest of efficiency, we require a white-box plannerthat directly adapts its search strategy during a planning cycle. We propose anefficient procedure for training adaptive search heuristics in a data-drivenimitation learning framework. We also draw a novel connection to Bayesianactive learning, and propose algorithms to adaptively evaluate edges of agraph. Our approach leads to the synthesis of a robust real-time planningmodule that allows a UAV to navigate seamlessly across environments andspeed-regimes. We evaluate our framework on a spectrum of planning problems andshow closed-loop results on 3 UAV platforms - a full-scale autonomoushelicopter, a large scale hexarotor and a small quadrotor. While the thesis wasmotivated by mobile robots, we have shown that the individual algorithms arebroadly applicable to other problem domains such as informative path planningand manipulation planning. We also establish novel connections between thedisparate fields of motion planning and active learning, imitation learning andonline paging which opens doors to several new research problems.

1 引言
2 專案背景
3 探索結構的規劃演算法
4 專家策劃的多元化組合
5 專家策劃的自適應組合
6 線上意外規劃器
7 基於模仿學習的資料驅動規劃
8 貝葉斯主動邊緣評估
9 統一的無人機規劃體系架構
10 結論
附錄A 無人機規劃問題
附錄B 動態投影濾波器
附錄C 無人機專家策劃庫
附錄D 黑盒規劃器的特徵提取

下載英文原文地址：

http://page5.dfpan.com/fs/2lc4j2221e29416ae17/

更多精彩文章請關注微訊號：